Professional Documents
Culture Documents
ISSN No:-2456-2165
Abstract:- This study examines the factors that influence Data analytics (DA) examines data sets to determine the
individual foreign direct investment decisions and predicts information they contain, increasingly with the support of spe-
using various Artificial Intelligence (A.I.) Algorithms mod- cialized systems and software.
els. The study also gives us an in-depth insight into the dy-
namics of complex FDI decisions using those A.I. predic- ARIMA models could generate projections for varieties
tive models. We use structural equation modeling in the of time series data. It is worth noting that the ARIMA model
prescriptive strand. Return on Investment (ROI), Secu- has three parts. However, not all aspects are always required,
rity/Personal Safety, and Investment Facilitation Services but it depends on the type of time-series dataset available. The
significantly affect individual FDI decisions. On predictive three parts are the autoregressive (A.R.), the integrated (I), and
strand analysis, we used various Machine Learning models finally, the moving average (M.A.). The assumption for the
to evaluate the accuracy of predicting classes of individual A.R. part of a time series dataset is that the observed value
FDI risk decisions and the ARIMA model for prediction. varies on some linear combinations of prior observed values
We find that Random Forest and Ada Boosting Trees have up to some upper limit lags plus an error term. The assumption
substantial classification accuracies despite the "No free for the M.A. part of time series data is that the practical esti-
lunch" theorem. The result also indicates that a better pre- mate is a random error term plus some linear permutations of
diction could be made by applying multiple classes of FDI previous random error terms up to some maximum lags
inflow decisions rather than binary classes. (Stanley Jere, 2017).
Keywords:- Foreign Direct Investment, Artificial Intelligence, The forecasts by the gradient boosting model and, of
Investment Facilitation, Return-On-Investment, Investment course, the random forest model is more precise than the target
Decisions, Predictive Modeling, Random Forest, Gradient forecasts. In comparison between the gradient boosting and
Boosting random forest models, the gradient boosting model turns out
to be more specific (Yoon, 2020)
I. INTRODUCTION
Generally, forecasting outcomes play a fundamental role
Foreign direct investment (FDI) is an investment in for policymakers, economic decision-makers, and other key
long‑term affiliation with a host country. It reflects a lasting stakeholders with the view of coming up with reasonable pol-
interest being controlled by a resident entity in one economy icies and suitable strategic plans but all those depend primarily
by a direct foreign stakeholder or a parent enterprise in an en- on the accuracy of the forecasts.
terprise resident or an economy different from that of the in-
vestor (United Nations, 2007). It should, however, be noted The main objective of this study is to conduct empirical
that foreign direct investment is different from indirect (port- research on the most effective A.I. application and propose ap-
folio) investment. FDI involves establishing a substantial, propriate recommendations to promote and facilitate invest-
long-term interest in the economy of a foreign country. ment in African Sub-Region based on predictive inference.
1
Exchange rate changes due to manipulation increase the risk and
uncertainty for investors
B. Analysis for Structural Equation Modelling (Descriptive) & Anderson, 2010). Bartlett's Test of Sphericity was signifi-
cant (Chi-square = 4924.914with df = 171, p.<0.001), imply-
Measurement model. ing significant intercorrelations. The five components had a
We conducted the case and variable screening (Yusuf, total variance of 71.245% (Hair et al., 2010). The total vari-
2020) and then analyzed the factor loadings and cross-loadings ance explained can be obtained in Table 1. For the final pattern
of the items. As a rule, items were maintained if (a) they had matrix that integrates Cronbach's alpha and the percentage of
high loadings on their primary factor. Typically, l >.30. And if variance explained, refer to Table 1
(b) they had minimal cross-loadings on some other factor (i.e.,
cross-loadings were less than half of their primary loadings We find that the factor Cronbach's alphas are robust;
(Hinkin, 1998). Kaiser-Meyer-Olkin (KMO) statistic was.659 Ease-of-Doing-Business is 0.959, Investment Promotion is
(Kaiser Meyer, 1970). The commonalities are acceptable; they 0.895, Social Desirability is 0.957, security is 0.841, and For-
are all above the minimum limit of 0.300 (Hair, Black, Babin, eign Exchange Rates is 0.804, respectively. The Correlation
2
This implies that the quantum of FDI inflow to a host country cou- of escalating the rates become higher which affect the investor’ de-
pled with finished manufactured products contribute to more influx cision to continue with investment or expansion.
of foreign currency into the host country which helps stabilize the
3
exchange rates, and on the other hand, if the rates are allowed to be FOREX rate is a high risk factor associated with investment uncer-
freely manipulated by the government of parallel market, then risk tainty, where sufficient exchange rate volatility usually increases ex-
pected profit.
TABLE 1: Pattern Matrix Table with Cronbach's Alpha Table 3: CFA Model Validity Analysis
K-Nearest Neighbor or K-NN is a Supervised Non-linear Fig 3: Error Rates by Trees number
classification algorithm. K-NN is a non-parametric algorithm; Result:
i.e. Algorithm; i.e., it doesn't assume underlying data or its dis-
tribution. It is one of the simplest and most widely used algo-
rithms, which depends on its k-value(Neighbors’-value)
The linear output variable is set to FALSE, considering A. Hypothetical Model Results for SEM
the impact of the independent variables on the dependent var- Our findings indicate that the hypotheses; Return on In-
iable is assumed to be non-linear. I also set the threshold to vestment has a positive effect on FDI inflow
0.01, connoting that if the change in error during an iteration (=P) (H1c), and Security/Personal Safety
was less than 1%, then no additional optimization would be Perception has a negative influence on FDI inflow (=
carried out by the model. Applying a (2,1) configuration 0.0187**) (H1). And Investment Facilitation Percep-
yielded 0.629% classification accuracy for this model tion has a positive effect on FDI inflow (90%, =
These are supported with significant P-
We then generated the error of the neural network model, Values <0.05. Nevertheless, the other theories (H1a, H1b,
along with the weights between the inputs, hidden layers, and H1d, H1f, and H1g) are not supported. Even though the meas-
output. We now predicted the rating using the neural network urements were not statistically significant, they were positive.
model. The expected rating would be scaled and transformed
to make a logical comparison with the actual rating. Sequel to These results also suggest the strongest (95% Confidence
that, we created a confusion matrix to compare the number of Interval) meaningful relationships among the Variable quanti-
true/false positives and negatives. ties of ROI on FDI inflow decisions (95%, = 0.9819), Se-
curity on FDI inflow decisions (95%, = 0.4902), and Invest-
Gradient Boosting ment Facilitation on FDI inflows (90%, = 1.0465). Refer to
Boosting is a sequential technique where each new model Figure 5 and Table 6 for results.
is built from learning the established errors of the previous
model, i.e., each predictor is trained using the residual errors Figure: 5: SEM Analysis Result
of the predecessor as labels.
The ANOVA test is performed, and the difference be- K-Means (KNN) Model
tween the null and residual deviance depicts how the model is
doing against the null model (a model with only the intercept). Fig 8: The graph indicates the proximity
The more significant this gap, the better it becomes. There is a
drop in deviance when we add each variable one at a time.
We also checked on the model misclassification. Mis- Table 8: Resampling minimal RMSE error
classification errors came out to be 41.93%. We could use re-
gression techniques with categorical variables to compare var-
ious other data. Although, as a rule of thumb, a model with a
sound predictive ability should have an AUC closer to 1 (1 is
ideal) than 0.5; this, and any other models, ranging from 0.52
to 0.63 because more than half of the dependent variables don't
have a significant P-value. We changed the parameters to im-
prove the AUC of this model from 0.5524 to 0.5767.
Neural Network
Given the output, we may conclude that both repetitions
converge. However, we will use the output driven in the sec-
ond repetition because it gives less error (97.40946) than the
error (100.43106) from the first repetition derives. We now
generate the error of the neural network model, along with the
weights between the inputs, hidden layers, and outputs, as
shown in Table 28 and Figure 16
Table 10: Error of the neural network Fig 12.: ROC-AUC for Gradient Boosting
Of the variables, FOREX, Security/safety, and Ease off- The AUC and Accuracy of the model are not that
Doing-Business are on top of the list strongly in tune with the model's algorithms and the predictors'
significance.
Decision Tree Model
Security has the highest variable importance, as depicted C. Consolidated Results Summaries
in Table 12.
Table 13: Consolidated summary of other ML models
Table 12: Split errors and Variable Importance
In model A, the ACF of the residuals shows no signifi- FDI decisions Prediction
cant autocorrelations, as shown in Appendix P. Even after con- Now we make a forecast at 95% CI and forecast for ten
version, the data remains stationary, with no difference after years using the ARIMA Model as depicted in Figures 16a and
being controlled. The Dickey-Fuller test returns a p-value of 16b.
0.8611, resulting in the rejection of the null hypothesis and ac-
cepting the alternative hypothesis that the data is stationary, as Fig 16a: Forecast for 10 Years
shown in Figure 15b Results: