The most fundamental driver of the BDI is the demand for commodities, which is influenced by factors such as global GDP, agricultural production and global energy demand (see Figure 2). The bulk shipping market is seasonal, cyclical and highly volatile (Zeng & Qu 2014). Because of the non-stationary and non-linear nature of price series and the complexity of influencing factors, it is difficult to analyse the fluctuations in the bulk shipping market. The relationship between GDP and the BDI is explained by assuming raw commodities such as corn, cocoa, coffee, primary metals and coal form the primary ingredients in industries such as the food, manufacturing and energy sectors (Feyrer 2009a, 2009b; Lin & Sim 2013; McPhail, Du & Muhammad 2012). Regional indexes like the Asian steel index have a major influence on the BDI (Chou et al. 2015) and individual changes in the larger economies like China that now accounts for 50% of iron ore imports (Lin & Sim 2013). The BDI price, for instance, decreased 38% in 2 months following the decision by the Chinese government to ban new investment in industries such as steel, automobiles and real estate (Hyung-Geun 2011).

The inelasticity of supply has a leverage effect on the BDI and supply inelasticity is a function of long lead times involved in new ship acquisitions that create a fixed tonne-mile capacity in the short term (Bakshi et al. 2011; Devanney 2010). The ability to change supply capacity is restricted to a 2–3-year period compared to the short-term fluctuations in the demand for cargo carriage (Koskinen & Hilmola 2005).

Other factors influencing the BDI index, include the lapse between the date the sea freight is contracted for and the date it is required (Laycan period)³ and the speculation effects by ship brokers. In addition, older vessels are more expensive to operate and this can also influence their respective hire rates (Alizadeh & Talley 2011). Bunker prices, closely linked to the price of crude oil, account for 25% – 33% of the total cost of operating a transport vessel. The volatility of crude oil prices, in turn, has an influence on the BDI price (Geman & Smith 2012; Notteboom 2012:#967).

The existence of bottlenecks (choke points) in global shipping routes has also been identified as a key factor driving the BDI price (Feyrer 2009a, Fu et al. 2010). Nearly half of the world’s oil tankers pass through a few narrow sea corridors, which are often subject to piracy, conflict, severe winters and accidents because of high traffic volumes. In recent times, the incidence of piracy in areas like the Somali coastline, the Gulf of Aden and the Suez Canal, has emphasised the risk (cost) associated with these choke points. The risk of piracy is illustrated in 2010 by 500 hostage incidents involving seafarers from more than 18 countries (Bowden et al. 2010). Although difficult to quantify, Bowden estimates that piracy costs the global economy between $7 and $12 billion per year. For some cargo operators, the increased risk associated with travel in these regions has rendered travel through the Gulf of Aden and the Suez Canal unfeasible. As a result, many cargo operators have decided to re-route via the Cape of Good Hope, which increases the distance by 3500 miles.

The length and severity of winters along key shipping routes, coupled with seasonal pressures such as reduced global crop yields during winter months, have been identified as further BDI price drivers. These pressures are especially significant along the choke points between January and March each year because of variations in factors like the extent of winter ice in the Baltic Sea (Koslowski & Loewe 1994).

Other factors influencing the BDI show that freight rates possess a 2.25–4.5-year cyclical tendency; it is supported by the fact that both the recessions of the 1970s and of the 1990s lasted approximately 5 years (Goulielmos & Psifia 2006). The mean reverting nature of freight rates, discussed above, appears to add credence to this argument (Tvedt 2003). If it is understood that freight rates tend to diverge from their mean value, and then revert to that value at some point in the future, it can be argued that this property is in itself a form of cyclicality, provided that the length of the reversion property is constant. In the case of freight rates, this is proved true through the findings of Goulielmos and Psifia (2006).

Changes in international shipping freight rates can predict international stock market returns (Alizadeh & Muradoglu 2014). A number of studies demonstrate that the BDI has a lagged relationship with share prices and commodities (OuYang, Wei & Zhang 2009). Empirical evidence suggests that this lag period appears to have varied between 1 and 3 months (Bakshi et al. 2011; Oomen 2012; OuYang et al. 2009). One of these studies, for example, that incorporated 23 developed and 25 undeveloped nations, showed a 1-month lag period between the BDI and global stock prices (Oomen 2012). Furthermore, it was found that the BDI was most effective as a predictor for global stock returns in the technology, telecommunications, consumer services and industrial sectors. In terms of the disaggregated BDI index, the Panamax appeared to be the best predictor of economic activity (Bakshi et al. 2011; Oomen 2012). Other research also demonstrated a significant relationship between the BDI and the MSCI Metals and Mining index (Kärrlander & Lanneström 2010) and that steel and corn prices show the strongest correlation to the BDI growth rate (Mariana 2008).

The data for the first research question largely consisted of narrative that was accessed from journal articles, maritime reports, financial and media commentary, as well as internet website articles, other reports and reviews. The data for the second question included two financial time-series data sets, namely the BDI and ALSI data indexes. Data from these two indexes were accessed from the Reuters and JSE websites, respectively, for the period May 1985–April 2016.

The first research question was to determine the fundamental dynamics of the BDI. Contents analysis incorporating axial and selective coding, was used to analyse the data. Firstly, axial analysis resulted in developing themes influencing the BDI. Selective coding then interrelated some of these and was used to develop an aggregated construct. Because of the multiplicity of factors influencing the dynamics of BDI, a qualitative approach, using an inductive lens, was thought to be the most suitable method (Cresswell & Plano Clark 2011; Ryan, Scapens & Theobald 2002).

The Zivot–Andrews unit root test was first employed to identify the break in the BDI data (Zivot & Andrews 2002). This endogenous structural break test incorporates the full sample and uses a range of different dummy variables for each potential break date. The actual break is selected where the t-statistic from the Augmented Dickey–Fuller (ADF) test of unit root is at its most negative value. Furthermore, the order of integration was calculated using the ADF unit root test (Dickey & Fuller 1979) with and without first differencing. The ADF test tests the null hypothesis that a unit root is present in a given time-series. The alternative hypothesis is usually stationarity or trend stationarity. The ADF test statistic is a negative number and the more negative it is, the stronger the rejection of the hypothesis that there is a unit root at a predefined level of confidence (i.e. usually 95%). The ADF test can handle more complex models than the standard Dickey–Fuller test, and it is more powerful (Harris 1992).

Correlation analysis for the two periods was then briefly inspected using Spearman’s rank correlation (based on Fisher’s transformation) between the two data sets. We employed the Johansen–Juselius test to ascertain whether the ALSI and BDI indexes were co-integrated. The Johansen–Juselius test is a multivariate generalisation of the ADF test and is an examination of linear combinations of variables for unit roots. The Johansen–Juselius test and estimation approach makes it possible to estimate all co-integrating vectors when there are more than two variables. If there are two variables (as in our case) each with unit roots, there can be at most one co-integrating vector (Johansen & Juselius 1990).

A Grainger causality model was constructed for both periods (Engle & Granger 1987) using different lag periods.

We tested for the absence of Granger causality by estimating the following VAR model: Yt=a0+a1Yt−1+…..+αpYt−p+b1Xt−1+…..+bpXt−p+ut[Eqn 1] Xt=c0+c1Xt−1+…..+cpXt−p+d1Yt−1+…..+dpYt−p+vt[Eqn 2]

Then, testing H₀: b₁ = b₂ = ….. = b_p = 0, against H_A: ‘Not H₀’, is a test that X does not Granger-cause Y. Similarly, testing H₀: d₁ = d₂ = ….. = d_p = 0, against H_A: ‘Not H₀’, is a test that Y does not Granger-cause X.

The properties of the model were tested at a range of lag periods (1–9 months). The optimum lag period was confirmed by the Akaike information criterion (AIC) (Akaike 1977). Finally, a range of diagnostics to ensure the validity of the Grainger causality model included the Wald test for Granger non-causality, where rejection of the null hypothesis implies causality. Furthermore, we employed the vector error-correction model (VECM) (Johansen 1988) to estimate the adjusted parameters (including an alpha to obtain the short-run adjustment parameters) for both ALSI and BDI to ascertain which one potentially responded more to shock or disequilibria.

Short- and long-run equilibria between ALSI and BDI were assessed using the VECM (Engle & Granger 1987) (second stage) with following specification: d(ALSI)=b2*d(BDI)+b3*εt−1+V[Eqn 3] where d(ALSI) is the first difference of ALSI, d(BDI) is the first difference of BDI, ε_t−1 is the one-period lag of residual obtained from the OLS estimation**, b2 is the short-run equilibrium coefficient and b3 is the coefficient for the one-period lag-correction term or residual from the first stage model and relates to the restoration of the system back to equilibrium, that is, long-run equilibrium and V is the error term.

**Ordinary least squares (OLS)-level model (first stage): ALSIt=b0+b1*BDIt+εt[Eqn 4] where ALSI is the All Share Index value, BDI is the Baltic Dry Index value, t is the month, ε_t is the error term (residual difference between observed and fitted values), b0 is the intercept and b1 is the slope coefficient for unlagged BDI versus ALSI.

Axial coding of the contents documents demonstrates a number of key themes that were demonstrated in all the literature (see Figure 3). These themes include the demand for shipping, the fixed nature of shipping supply, the Laycan period, vessel size and age, bunker prices, the effect of piracy and global winters. Other factors include the number of shipping routes (25), the number of vessels, port costs, administration and a cyclical component. Selective coding was used to aggregate the axial codes into four major constructs. This included a demand and supply function that largely incorporated global demand for shipping versus shipping capacity.

The other axial codes included a risk management and speculative component (Laycan period, broker speculation, piracy, climatic conditions and a cyclical component), a logistics construct (number of vessels, number of routes, choke points and climatic conditions) and a cost management aspect (cost and maintenance of vessels, bunker oil prices and docking fees). The BDI, which is disaggregated into four major components (see Figure 3), is thus influenced by a complex set of variables that make it difficult to model empirically.

The Zivot–Andrews unit root test was first employed to identify the break in the BDI data (Zivot & Andrews 2002) to effectively create two data sets. Correlation analysis was used to illustrate the association in both periods, before confirming the first data sets were co-integrated using the Johansen–Juselius test for co-integration. An optimal Granger causality model was then developed and tested. Lastly, a VECM was developed to assess the magnitude of the coefficients and also ascertain which variable was leading and which was lagging.

Figure 4 depicts the relationship between monthly BDI price and JSE ALSI price movements for the period 1985–April 2016. An application of the Zivot–Andrews test indicates a structural break in the BDI series, occurred in July 2008 (minimum t-statistic was −5.641: critical values: 1%: −5.57, 5%: −5.08, 10%: −4.82). From 1985 to July 2008, the BDI had a steady upward trend that appeared consistent with the JSE ALSI trend. After July 2008, the BDI declines markedly, while the ALSI index maintains its upward trend to 2016.

The MacKinnon approximate p-values from the ADF test for unit without differencing for both ALSI and BDI with a trend component were 0.852 and 0.218, respectively (see Table 1). As a trend in levels becomes a constant in first differences, no trend was included for the first differenced Dickey–Fuller tests but drift assumed. Both null hypotheses are rejected for the differenced series (p < 0.01 for both).

As both raw series are not stationary but the first difference is stationary and invertible as suggested by the ADF test results, this suggests that both times series are (1).

In the first period up to June 2008, the correlation = 0.836 on 138 observations (95% confidence interval (CI): 0.797–0.868), as illustrated in Figure 5. It is clear in this period that both the BDI and the ALSI have increased, showing that they were positively correlated (red line) with a linear trend between the two indexes (green line) to reflect both a higher BDI price as well as increased share prices.

In the second period (July 2008–April 2016), we observed a negative correlation of −0.677 on (95% CI: 0.774 to −0.550) clearly illustrating the downward turn in the BDI index, while the ALSI continued to increase.

Prior to the break in 2008, the first hypothesis for the Johansen–Juselius test, r = 0, which tests for the presence of co-integration, had a test statistic of 23.21 (which exceeded the 17.95 critical cut-off at 5% level and was close to the 1% critical value of 23.52), that is, p < 0.05, and thus suggested that we have evidence to reject the null hypothesis of no co-integration, that is, the two series (ALSI and BDI) were co-integrated prior to the break. We observed a similar finding post the identified structural break in 2008, with a test statistic of 23.21 for r = 0 (which exceeded the 23.52 critical cut-off at 1% level), that is, p < 0.01, and thus suggested that we have strong evidence to reject the null hypothesis of no co-integration, that is, the two series (ALSI and BDI) were also co-integrated post-break.

The results of the Granger causality test pre-break (January 1985–June 2008) indicated a p-value of 0.001), and thus we rejected the null hypothesis of no Granger causality (Table 2). This indicates that Model 2 (ALSI ~ Lags(ALSI, 1:6), i.e. without BDI as a causal covariate) is too restrictive as compared with Model 1 (Model 1: ALSI ~ Lags(ALSI, 1:6) + Lags(BDI, 1:6), i.e. with BDI included as a covariate). This relationship was no longer significant post-break (p = 0.246), that is, for the post-break period, we observed an inverse and non-linear relationship between BDI and ALSI.

The adjusted parameter estimates for ALSI and BDI from the VECM model suggest that ALSI may respond more if there is change or shock in the system compared to BDI [alpha coefficient (α) = 0.0032 (p = 0.017) and α = 0.0016 (p = 0.025)], respectively. However, the overlap of the confidence intervals (CIs) suggests that these alphas are not significantly different (α 95% CIs of 0.0006–0.0057 and 0.0002–0.003, respectively).

The findings, based on an inductive review of a wide range of data, suggest that four key constructs influenced the BDI price between 1985 and 2009. The validity of the themes (axial codes) in our results are underpinned by a wide range of literature that suggests that the BDI is influenced by the inelasticity of shipping supply volume (Koskinen & Hilmola 2005), global commodity demand (Oomen 2012), country-level effects like the ‘Chinese Factor’ (Hyung-Geun 2011) and the contract details and speculation effect (Laycan period) between ship owners and charterers (Alizadeh & Talley 2011). Other factors included costs such as bunker prices, the maintenance of vessels and other freight costs (Notteboom & Vernimmen 2009), and choke points in global shipping routes because of issues like weather (the Baltic Winter) and the effects of piracy (Fu et al. 2010). Aggregating these themes, this article proposes that they are better understood in terms of four basic constructs, namely supply and demand of dry bulk shipping, and a combination of risk, cost and logistics management factors.

The BDI is currently at an all-time low. Recent speculation indicates that the dynamics of the BDI have changed because of an oversupply of shipping capacity. Increased demand for dry bulk cargo is therefore unlikely to be reflected in an increase in the BDI in the short term because of this excess capacity, which was commissioned when the BDI was at a peak in 2009/2010. A question therefore exists as to whether the historic dynamics of the BDI will resume over the longer term as supply and capacity are eventually matched.